Skip to main content

Performance & mobile troubleshooting

Performance problems can come from the browser/client, conversation size, network streaming, model latency, or the Starfire backend. Narrow the layer before changing model or account settings.

App feels slow only in one long chat

Very large rendered histories can increase browser memory and layout work, especially on mobile. Try:
  • start a new conversation for a new task
  • move durable context into Projects/Knowledge
  • avoid keeping unnecessary large attachments in one active workflow
  • reload the client after preserving any unsent work

Response begins slowly

Separate model/provider latency from UI rendering latency. If the request ID shows the server started late, investigate route/provider health. If output arrives but the UI renders poorly, focus on client performance.

Streaming stutters

Possible causes include:
  • unstable network connection
  • large UI tree/history
  • browser/device resource pressure
  • provider stream interruption
  • client reconnect behavior
Test a short new conversation to determine whether the issue follows the model or the existing conversation UI.

Mobile navigation issue

If a sidebar, modal, or navigation drawer becomes stuck:
  • reload the current page
  • confirm the issue reproduces after a fresh session
  • capture device/client version and page path
  • test another route to determine whether the bug is page-specific

Browser-specific issue

Compare another supported browser/client when practical. If only one client fails, capture its version and console-visible error/request ID where safe.

Platform-wide slowness

If multiple users and unrelated workflows become slow simultaneously, operators should review API latency, providers, database/cache, queues/workers, and incidents.

What to include in a report

  • device/client/browser
  • page or feature
  • conversation/build/run ID when applicable
  • request ID
  • approximate timestamp
  • whether a fresh conversation reproduces it
  • whether another model/client behaves differently
A mobile performance problem can be real even when the model backend is healthy. Treat frontend rendering and AI execution as separate layers during diagnosis.